Supporting case
NVIDIA releases a smaller Bolt model for local assistants
What happened
NVIDIA’s model card dates the Nemotron 3.5 Bolt E4B release October 23. Downloadable quantized weights and an OpenMDW1.1 licence accompany a model intended for local assistants and embedded interfaces. It has eight billion total parameters and four billion “effective” parameters; the smaller figure is not the full stored model.
The card includes local-serving instructions but also labels the llama.cpp deployment recipe pending. Its “4GB” name should not be read as independently verified total memory requirements for a complete application. Benchmark results are NVIDIA’s measurements. Model card and files ↗.
Downloadable smaller weights could widen independent deployment. Usable capability still depends on hardware, software integration and developer-measured performance.
Dates and implementation stage
Event 2026-10-23; approved October 25; Publication dates and unresolved timestamp qualifications remain in the account
